Window Seat LoRA for Wan 2.2: Turn Travel Views Into Cinematic Surrealism

TL;DR:

• Start every prompt with the trigger phrase "Window Seat", then describe a view through a vehicle window.
• Place [WAN2.2]WindowSeat_Redmond_high_noise.safetensors in ComfyUI/models/loras/ and use a LoRA strength of 0.8–1.0.
• Built for Wan-AI/Wan2.2-T2V-A14B, released under CC0, trained with GPU support from Redmond AI.

What the Window Seat LoRA Does

WindowSeat is a text-to-video LoRA for the Wan 2.2 T2V A14B base model. According to the HuggingFace model card, it captures "the contemplative, surreal experience of looking through windows while traveling." Every scene the model produces is a view from a bus, train, plane, subway, or tram that reveals an extraordinary moment beyond ordinary reality.

The model card's example prompts show the range: a one-room schoolhouse isolated in a sunflower field seen from a moving bus, a desert turning into mirror-like salt flats below a plane during turbulence, floating islands connected by rope-and-lantern bridges above the clouds, a spotlit giraffe on an empty theater stage seen from a subway, and an orchestra of cellists performing in a wheat field under stormy skies.

How It Was Trained

The LoRA is built on Wan-AI/Wan2.2-T2V-A14B, Wan's open text-to-video 14B model. Training was sponsored by Redmond AI, which provided GPU time, according to the model card's special thanks note. The released file is [WAN2.2]WindowSeat_Redmond_high_noise.safetensors (458.69 MB, autoV2 hash E8EB3B532C on Civitai). The model is licensed CC0 1.0.

Trigger Words and Prompting

The trigger phrase is "Window Seat" — one two-word phrase, listed as the trained word on both the Civitai page and the model card. The author recommends starting your prompt with it, followed by a description of the surreal view.

The example prompts on the model card are structured, cinematic descriptions rather than tag lists. A typical pattern:

Window Seat. A locked-off shot from the interior of a moving subway captures the window view of an old theater stage where a single spotlight illuminates a giraffe standing center stage.

They pair the shot type and camera movement with the surreal scene and a finishing style cue such as "warm 16mm film grain," "cinematic grain," or "Tarkovsky's contemplative surrealism."

Setup for ComfyUI

1. Download [WAN2.2]WindowSeat_Redmond_high_noise.safetensors and place it in ComfyUI/models/loras/.
2. Load your base Wan model.
3. Add a Load LoRA node connected to the model loader output.
4. Select the WindowSeat safetensors file in the LoRA node.
5. Connect the MODEL output of the LoRA node to your KSampler.
6. Set strength_model between 0.8 and 1.0 as a starting point.

Your ComfyUI installation must already be configured to run Wan models. The model card also lists the diffusers pipeline tag on HuggingFace, so the LoRA can be loaded with the standard LoraLoaderMixin flow in diffusers.

What the Samples Actually Show

Frames extracted from the model card's sample videos show a consistent look: a cinematic point of view from inside the vehicle, with the window framing the scene, heavy atmospheric grading, and film grain. Subjects match the prompts — salt flats reflecting storm clouds, floating islands over a cloud layer, and other surreal scenes — with painterly, contemplative framing rather than photoreal sharpness.

Common Errors and Fixes

Error | Cause | Fix | Source

|---|---|---|---

Model has no effect on output | Trigger phrase missing | Start the prompt with "Window Seat" | HF model card LoRA not found in ComfyUI | File in wrong folder | Place it in ComfyUI/models/loras/ | HF model card Generated shots ignore the window framing | Strength too low or scene too plain | Raise strength_model to 0.8–1.0; describe a view through a vehicle window | HF model card

FAQ

Which base model does this LoRA require?

Wan-AI/Wan2.2-T2V-A14B, per the base_model field in the HuggingFace model card metadata.

Is it free to use?

The model is released under CC0 1.0.

What is the trigger word?

"Window Seat" — the Civitai trained-words list and the HF model card both give this exact phrase.

Can I use it with diffusers?

The HuggingFace model card is tagged with the diffusers library and text-to-video pipeline, and the LoRA is built on a supported base model.

Sources

• HuggingFace model card: WindowSeat-Redmond-WAN2-T2V-14B — description, trigger phrase, ComfyUI setup, example prompts, CC0 license.
• [Civitai: [WAN2.2] WindowSeat - Redmond - T2V-14B - v1.0](https://civitai.com/models/2071196) — model stats, base model, trained words, file hash.
• Wan-AI/Wan2.2-T2V-A14B — the base model this LoRA adapts.